Meta Open‑Sources BOxCrete
What happened
Meta published BOxCrete, an AI model that uses Bayesian optimization to design concrete mixes for data‑center construction, reporting 43% faster strength gains and 10% less cracking using domestic materials. The release shows how AI is being applied to engineering operations, not just consumer features, and Meta provided a technical blog for deeper details. (x.com)
Why it matters
Meta published BOxCrete and announced the release alongside the American Concrete Institute Spring Convention on March 30, 2026, and made the code and foundational dataset available as open source on GitHub. (engineering.fb.com)(github.com) The system was used in collaboration with ready‑mix supplier Amrize, contractor Mortenson, and researchers at the University of Illinois to produce a specialized mix for a site‑support slab at Meta’s Rosemount, Minnesota data center; the mix performance met metrics that qualified it for additional areas of the facility. (enr.com)(amrize.com) BOxCrete applies Bayesian optimization — a statistical search method that picks the next candidate mixtures by weighing expected improvement against uncertainty — on top of a probabilistic surrogate model so the system can suggest mixes that are likely both high-performing and informative. (arxiv.org)(github.com) The surrogate model in BOxCrete is a Gaussian Process, which means it learns a smooth function mapping mix ingredients to strength while also giving an uncertainty estimate for each prediction; the authors report a predictive fit (R²) of about 0.94, meaning the model explains roughly 94% of observed variance, and a root‑mean‑square error of 0.69 ksi, meaning average prediction errors are about 0.69 thousand pounds per square inch. (arxiv.org) The project’s dataset combines over 500 compressive‑strength measurements from 123 mixes tested at explicit curing ages (1, 3, 5, 14, and 28 days), and the optimization extends beyond peak strength to include workability (slump, which is a laboratory measure of how fluid and placeable fresh concrete is) and embodied carbon (the mix’s global warming potential from its ingredients). (arxiv.org)(github.com) The released repository includes code, notebooks, and data under an MIT license so practitioners and researchers can reproduce the training, run the optimization workflows, and adapt the approach to different local materials or project constraints; the technical paper describing BOxCrete was posted to arXiv on March 23, 2026. (github.com)(arxiv.org)
Key numbers
- Meta published BOxCrete, an AI model that uses Bayesian optimization to design concrete mixes for data‑center construction, reporting 43% faster strength gains and 10% less cracking using domestic materials.
- (x.com) Meta published BOxCrete and announced the release alongside the American Concrete Institute Spring Convention on March 30, 2026, and made the code and foundational dataset available as open source on GitHub.
Quick answers
What happened in Meta Open‑Sources BOxCrete?
Meta published BOxCrete, an AI model that uses Bayesian optimization to design concrete mixes for data‑center construction, reporting 43% faster strength gains and 10% less cracking using domestic materials. The release shows how AI is being applied to engineering operations, not just consumer features, and Meta provided a technical blog for deeper details. (x.com)
Why does Meta Open‑Sources BOxCrete matter?
Meta published BOxCrete and announced the release alongside the American Concrete Institute Spring Convention on March 30, 2026, and made the code and foundational dataset available as open source on GitHub. (engineering.fb.com)(github.com) The system was used in collaboration with ready‑mix supplier Amrize, contractor Mortenson, and researchers at the University of Illinois to produce a specialized mix for a site‑support slab at Meta’s Rosemount, Minnesota data center; the mix performance met metrics that qualified it for additional areas of the facility. (enr.com)(amrize.com) BOxCrete applies Bayesian optimization — a statistical search method that picks the next candidate mixtures by weighing expected improvement against uncertainty — on top of a probabilistic surrogate model so the system can suggest mixes that are likely both high-performing and informative. (arxiv.org)(github.com) The surrogate model in BOxCrete is a Gaussian Process, which means it learns a smooth function mapping mix ingredients to strength while also giving an uncertainty estimate for each prediction; the authors report a predictive fit (R²) of about 0.94, meaning the model explains roughly 94% of observed variance, and a root‑mean‑square error of 0.69 ksi, meaning average prediction errors are about 0.69 thousand pounds per square inch. (arxiv.org) The project’s dataset combines over 500 compressive‑strength measurements from 123 mixes tested at explicit curing ages (1, 3, 5, 14, and 28 days), and the optimization extends beyond peak strength to include workability (slump, which is a laboratory measure of how fluid and placeable fresh concrete is) and embodied carbon (the mix’s global warming potential from its ingredients). (arxiv.org)(github.com) The released repository includes code, notebooks, and data under an MIT license so practitioners and researchers can reproduce the training, run the optimization workflows, and adapt the approach to different local materials or project constraints; the technical paper describing BOxCrete was posted to arXiv on March 23, 2026. (github.com)(arxiv.org)
